Long-context retrieval and reasoning on RULER V3
100NIAH Single Query 1Full Prompt
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Full PromptTarget Model=DeepSeek V3, Keep rate=-2026.03 | 100 | 100 | 62.8 | 89.2 | 99.6 | 32.2 | 98.75 | 99.25 | 61.2 | 55 | 80 | |
| Spec prefill with Qwen3-4B-InstructTarget Model=DeepSeek V3, Draft Model=Qwen3-4B-Instruct, Keep rate=12.5%2026.03 | 100 | 100 | 100 | 100 | 83.4 | 76.4 | 99.6 | 100 | 70.46 | 66.84 | 89.67 | |
| Spec prefill with Llama-3.1-8B-InstructTarget Model=DeepSeek V3, Draft Model=Llama-3.1-8B-Instruct, Keep rate=12.5%2026.03 | 100 | 100 | 100 | 99.8 | 99.6 | 96.8 | 99.2 | 100 | 70.74 | 67.43 | 93.36 |